tools/perf/builtin-diff.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * builtin-diff.c
   4  *
   5  * Builtin diff command: Analyze two perf.data input files, look up and read
   6  * DSOs and symbol information, sort them and produce a diff.
   7  */
   8 #include "builtin.h"
   9
  10 #include "util/debug.h"
  11 #include "util/event.h"
  12 #include "util/hist.h"
  13 #include "util/evsel.h"
  14 #include "util/evlist.h"
  15 #include "util/session.h"
  16 #include "util/tool.h"
  17 #include "util/sort.h"
  18 #include "util/srcline.h"
  19 #include "util/symbol.h"
  20 #include "util/data.h"
  21 #include "util/config.h"
  22 #include "util/time-utils.h"
  23 #include "util/annotate.h"
  24 #include "util/map.h"
  25 #include "util/spark.h"
  26 #include "util/block-info.h"
  27 #include "util/stream.h"
  28 #include "util/util.h"
  29 #include <linux/err.h>
  30 #include <linux/zalloc.h>
  31 #include <subcmd/pager.h>
  32 #include <subcmd/parse-options.h>
  33
  34 #include <errno.h>
  35 #include <inttypes.h>
  36 #include <stdlib.h>
  37 #include <math.h>
  38
  39 struct perf_diff {
  40         struct perf_tool                 tool;
  41         const char                      *time_str;
  42         struct perf_time_interval       *ptime_range;
  43         int                              range_size;
  44         int                              range_num;
  45         bool                             has_br_stack;
  46         bool                             stream;
  47 };
  48
  49 /* Diff command specific HPP columns. */
  50 enum {
  51         PERF_HPP_DIFF__BASELINE,
  52         PERF_HPP_DIFF__PERIOD,
  53         PERF_HPP_DIFF__PERIOD_BASELINE,
  54         PERF_HPP_DIFF__DELTA,
  55         PERF_HPP_DIFF__RATIO,
  56         PERF_HPP_DIFF__WEIGHTED_DIFF,
  57         PERF_HPP_DIFF__FORMULA,
  58         PERF_HPP_DIFF__DELTA_ABS,
  59         PERF_HPP_DIFF__CYCLES,
  60         PERF_HPP_DIFF__CYCLES_HIST,
  61
  62         PERF_HPP_DIFF__MAX_INDEX
  63 };
  64
  65 struct diff_hpp_fmt {
  66         struct perf_hpp_fmt      fmt;
  67         int                      idx;
  68         char                    *header;
  69         int                      header_width;
  70 };
  71
  72 struct data__file {
  73         struct perf_session     *session;
  74         struct perf_data         data;
  75         int                      idx;
  76         struct hists            *hists;
  77         struct evlist_streams   *evlist_streams;
  78         struct diff_hpp_fmt      fmt[PERF_HPP_DIFF__MAX_INDEX];
  79 };
  80
  81 static struct data__file *data__files;
  82 static int data__files_cnt;
  83
  84 #define data__for_each_file_start(i, d, s)      \
  85         for (i = s, d = &data__files[s];        \
  86              i < data__files_cnt;               \
  87              i++, d = &data__files[i])
  88
  89 #define data__for_each_file(i, d) data__for_each_file_start(i, d, 0)
  90 #define data__for_each_file_new(i, d) data__for_each_file_start(i, d, 1)
  91
  92 static bool force;
  93 static bool show_period;
  94 static bool show_formula;
  95 static bool show_baseline_only;
  96 static bool cycles_hist;
  97 static unsigned int sort_compute = 1;
  98
  99 static s64 compute_wdiff_w1;
 100 static s64 compute_wdiff_w2;
 101
 102 static const char               *cpu_list;
 103 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
 104
 105 enum {
 106         COMPUTE_DELTA,
 107         COMPUTE_RATIO,
 108         COMPUTE_WEIGHTED_DIFF,
 109         COMPUTE_DELTA_ABS,
 110         COMPUTE_CYCLES,
 111         COMPUTE_MAX,
 112         COMPUTE_STREAM, /* After COMPUTE_MAX to avoid use current compute arrays */
 113 };
 114
 115 const char *compute_names[COMPUTE_MAX] = {
 116         [COMPUTE_DELTA] = "delta",
 117         [COMPUTE_DELTA_ABS] = "delta-abs",
 118         [COMPUTE_RATIO] = "ratio",
 119         [COMPUTE_WEIGHTED_DIFF] = "wdiff",
 120         [COMPUTE_CYCLES] = "cycles",
 121 };
 122
 123 static int compute = COMPUTE_DELTA_ABS;
 124
 125 static int compute_2_hpp[COMPUTE_MAX] = {
 126         [COMPUTE_DELTA]         = PERF_HPP_DIFF__DELTA,
 127         [COMPUTE_DELTA_ABS]     = PERF_HPP_DIFF__DELTA_ABS,
 128         [COMPUTE_RATIO]         = PERF_HPP_DIFF__RATIO,
 129         [COMPUTE_WEIGHTED_DIFF] = PERF_HPP_DIFF__WEIGHTED_DIFF,
 130         [COMPUTE_CYCLES]        = PERF_HPP_DIFF__CYCLES,
 131 };
 132
 133 #define MAX_COL_WIDTH 70
 134
 135 static struct header_column {
 136         const char *name;
 137         int width;
 138 } columns[PERF_HPP_DIFF__MAX_INDEX] = {
 139         [PERF_HPP_DIFF__BASELINE] = {
 140                 .name  = "Baseline",
 141         },
 142         [PERF_HPP_DIFF__PERIOD] = {
 143                 .name  = "Period",
 144                 .width = 14,
 145         },
 146         [PERF_HPP_DIFF__PERIOD_BASELINE] = {
 147                 .name  = "Base period",
 148                 .width = 14,
 149         },
 150         [PERF_HPP_DIFF__DELTA] = {
 151                 .name  = "Delta",
 152                 .width = 7,
 153         },
 154         [PERF_HPP_DIFF__DELTA_ABS] = {
 155                 .name  = "Delta Abs",
 156                 .width = 7,
 157         },
 158         [PERF_HPP_DIFF__RATIO] = {
 159                 .name  = "Ratio",
 160                 .width = 14,
 161         },
 162         [PERF_HPP_DIFF__WEIGHTED_DIFF] = {
 163                 .name  = "Weighted diff",
 164                 .width = 14,
 165         },
 166         [PERF_HPP_DIFF__FORMULA] = {
 167                 .name  = "Formula",
 168                 .width = MAX_COL_WIDTH,
 169         },
 170         [PERF_HPP_DIFF__CYCLES] = {
 171                 .name  = "[Program Block Range] Cycles Diff",
 172                 .width = 70,
 173         },
 174         [PERF_HPP_DIFF__CYCLES_HIST] = {
 175                 .name  = "stddev/Hist",
 176                 .width = NUM_SPARKS + 9,
 177         }
 178 };
 179
 180 static int setup_compute_opt_wdiff(char *opt)
 181 {
 182         char *w1_str = opt;
 183         char *w2_str;
 184
 185         int ret = -EINVAL;
 186
 187         if (!opt)
 188                 goto out;
 189
 190         w2_str = strchr(opt, ',');
 191         if (!w2_str)
 192                 goto out;
 193
 194         *w2_str++ = 0x0;
 195         if (!*w2_str)
 196                 goto out;
 197
 198         compute_wdiff_w1 = strtol(w1_str, NULL, 10);
 199         compute_wdiff_w2 = strtol(w2_str, NULL, 10);
 200
 201         if (!compute_wdiff_w1 || !compute_wdiff_w2)
 202                 goto out;
 203
 204         pr_debug("compute wdiff w1(%" PRId64 ") w2(%" PRId64 ")\n",
 205                   compute_wdiff_w1, compute_wdiff_w2);
 206
 207         ret = 0;
 208
 209  out:
 210         if (ret)
 211                 pr_err("Failed: wrong weight data, use 'wdiff:w1,w2'\n");
 212
 213         return ret;
 214 }
 215
 216 static int setup_compute_opt(char *opt)
 217 {
 218         if (compute == COMPUTE_WEIGHTED_DIFF)
 219                 return setup_compute_opt_wdiff(opt);
 220
 221         if (opt) {
 222                 pr_err("Failed: extra option specified '%s'", opt);
 223                 return -EINVAL;
 224         }
 225
 226         return 0;
 227 }
 228
 229 static int setup_compute(const struct option *opt, const char *str,
 230                          int unset __maybe_unused)
 231 {
 232         int *cp = (int *) opt->value;
 233         char *cstr = (char *) str;
 234         char buf[50];
 235         unsigned i;
 236         char *option;
 237
 238         if (!str) {
 239                 *cp = COMPUTE_DELTA;
 240                 return 0;
 241         }
 242
 243         option = strchr(str, ':');
 244         if (option) {
 245                 unsigned len = option++ - str;
 246
 247                 /*
 248                  * The str data are not writeable, so we need
 249                  * to use another buffer.
 250                  */
 251
 252                 /* No option value is longer. */
 253                 if (len >= sizeof(buf))
 254                         return -EINVAL;
 255
 256                 strncpy(buf, str, len);
 257                 buf[len] = 0x0;
 258                 cstr = buf;
 259         }
 260
 261         for (i = 0; i < COMPUTE_MAX; i++)
 262                 if (!strcmp(cstr, compute_names[i])) {
 263                         *cp = i;
 264                         return setup_compute_opt(option);
 265                 }
 266
 267         pr_err("Failed: '%s' is not computation method "
 268                "(use 'delta','ratio' or 'wdiff')\n", str);
 269         return -EINVAL;
 270 }
 271
 272 static double period_percent(struct hist_entry *he, u64 period)
 273 {
 274         u64 total = hists__total_period(he->hists);
 275
 276         return (period * 100.0) / total;
 277 }
 278
 279 static double compute_delta(struct hist_entry *he, struct hist_entry *pair)
 280 {
 281         double old_percent = period_percent(he, he->stat.period);
 282         double new_percent = period_percent(pair, pair->stat.period);
 283
 284         pair->diff.period_ratio_delta = new_percent - old_percent;
 285         pair->diff.computed = true;
 286         return pair->diff.period_ratio_delta;
 287 }
 288
 289 static double compute_ratio(struct hist_entry *he, struct hist_entry *pair)
 290 {
 291         double old_period = he->stat.period ?: 1;
 292         double new_period = pair->stat.period;
 293
 294         pair->diff.computed = true;
 295         pair->diff.period_ratio = new_period / old_period;
 296         return pair->diff.period_ratio;
 297 }
 298
 299 static s64 compute_wdiff(struct hist_entry *he, struct hist_entry *pair)
 300 {
 301         u64 old_period = he->stat.period;
 302         u64 new_period = pair->stat.period;
 303
 304         pair->diff.computed = true;
 305         pair->diff.wdiff = new_period * compute_wdiff_w2 -
 306                            old_period * compute_wdiff_w1;
 307
 308         return pair->diff.wdiff;
 309 }
 310
 311 static int formula_delta(struct hist_entry *he, struct hist_entry *pair,
 312                          char *buf, size_t size)
 313 {
 314         u64 he_total = he->hists->stats.total_period;
 315         u64 pair_total = pair->hists->stats.total_period;
 316
 317         if (symbol_conf.filter_relative) {
 318                 he_total = he->hists->stats.total_non_filtered_period;
 319                 pair_total = pair->hists->stats.total_non_filtered_period;
 320         }
 321         return scnprintf(buf, size,
 322                          "(%" PRIu64 " * 100 / %" PRIu64 ") - "
 323                          "(%" PRIu64 " * 100 / %" PRIu64 ")",
 324                          pair->stat.period, pair_total,
 325                          he->stat.period, he_total);
 326 }
 327
 328 static int formula_ratio(struct hist_entry *he, struct hist_entry *pair,
 329                          char *buf, size_t size)
 330 {
 331         double old_period = he->stat.period;
 332         double new_period = pair->stat.period;
 333
 334         return scnprintf(buf, size, "%.0F / %.0F", new_period, old_period);
 335 }
 336
 337 static int formula_wdiff(struct hist_entry *he, struct hist_entry *pair,
 338                          char *buf, size_t size)
 339 {
 340         u64 old_period = he->stat.period;
 341         u64 new_period = pair->stat.period;
 342
 343         return scnprintf(buf, size,
 344                   "(%" PRIu64 " * " "%" PRId64 ") - (%" PRIu64 " * " "%" PRId64 ")",
 345                   new_period, compute_wdiff_w2, old_period, compute_wdiff_w1);
 346 }
 347
 348 static int formula_fprintf(struct hist_entry *he, struct hist_entry *pair,
 349                            char *buf, size_t size)
 350 {
 351         switch (compute) {
 352         case COMPUTE_DELTA:
 353         case COMPUTE_DELTA_ABS:
 354                 return formula_delta(he, pair, buf, size);
 355         case COMPUTE_RATIO:
 356                 return formula_ratio(he, pair, buf, size);
 357         case COMPUTE_WEIGHTED_DIFF:
 358                 return formula_wdiff(he, pair, buf, size);
 359         default:
 360                 BUG_ON(1);
 361         }
 362
 363         return -1;
 364 }
 365
 366 static void *block_hist_zalloc(size_t size)
 367 {
 368         struct block_hist *bh;
 369
 370         bh = zalloc(size + sizeof(*bh));
 371         if (!bh)
 372                 return NULL;
 373
 374         return &bh->he;
 375 }
 376
 377 static void block_hist_free(void *he)
 378 {
 379         struct block_hist *bh;
 380
 381         bh = container_of(he, struct block_hist, he);
 382         hists__delete_entries(&bh->block_hists);
 383         free(bh);
 384 }
 385
 386 struct hist_entry_ops block_hist_ops = {
 387         .new    = block_hist_zalloc,
 388         .free   = block_hist_free,
 389 };
 390
 391 static int diff__process_sample_event(struct perf_tool *tool,
 392                                       union perf_event *event,
 393                                       struct perf_sample *sample,
 394                                       struct evsel *evsel,
 395                                       struct machine *machine)
 396 {
 397         struct perf_diff *pdiff = container_of(tool, struct perf_diff, tool);
 398         struct addr_location al;
 399         struct hists *hists = evsel__hists(evsel);
 400         struct hist_entry_iter iter = {
 401                 .evsel  = evsel,
 402                 .sample = sample,
 403                 .ops    = &hist_iter_normal,
 404         };
 405         int ret = -1;
 406
 407         if (perf_time__ranges_skip_sample(pdiff->ptime_range, pdiff->range_num,
 408                                           sample->time)) {
 409                 return 0;
 410         }
 411
 412         addr_location__init(&al);
 413         if (machine__resolve(machine, &al, sample) < 0) {
 414                 pr_warning("problem processing %d event, skipping it.\n",
 415                            event->header.type);
 416                 ret = -1;
 417                 goto out;
 418         }
 419
 420         if (cpu_list && !test_bit(sample->cpu, cpu_bitmap)) {
 421                 ret = 0;
 422                 goto out;
 423         }
 424
 425         switch (compute) {
 426         case COMPUTE_CYCLES:
 427                 if (!hists__add_entry_ops(hists, &block_hist_ops, &al, NULL,
 428                                           NULL, NULL, NULL, sample, true)) {
 429                         pr_warning("problem incrementing symbol period, "
 430                                    "skipping event\n");
 431                         goto out;
 432                 }
 433
 434                 hist__account_cycles(sample->branch_stack, &al, sample, false,
 435                                      NULL);
 436                 break;
 437
 438         case COMPUTE_STREAM:
 439                 if (hist_entry_iter__add(&iter, &al, PERF_MAX_STACK_DEPTH,
 440                                          NULL)) {
 441                         pr_debug("problem adding hist entry, skipping event\n");
 442                         goto out;
 443                 }
 444                 break;
 445
 446         default:
 447                 if (!hists__add_entry(hists, &al, NULL, NULL, NULL, NULL, sample,
 448                                       true)) {
 449                         pr_warning("problem incrementing symbol period, "
 450                                    "skipping event\n");
 451                         goto out;
 452                 }
 453         }
 454
 455         /*
 456          * The total_period is updated here before going to the output
 457          * tree since normally only the baseline hists will call
 458          * hists__output_resort() and precompute needs the total
 459          * period in order to sort entries by percentage delta.
 460          */
 461         hists->stats.total_period += sample->period;
 462         if (!al.filtered)
 463                 hists->stats.total_non_filtered_period += sample->period;
 464         ret = 0;
 465 out:
 466         addr_location__exit(&al);
 467         return ret;
 468 }
 469
 470 static struct perf_diff pdiff = {
 471         .tool = {
 472                 .sample = diff__process_sample_event,
 473                 .mmap   = perf_event__process_mmap,
 474                 .mmap2  = perf_event__process_mmap2,
 475                 .comm   = perf_event__process_comm,
 476                 .exit   = perf_event__process_exit,
 477                 .fork   = perf_event__process_fork,
 478                 .lost   = perf_event__process_lost,
 479                 .namespaces = perf_event__process_namespaces,
 480                 .cgroup = perf_event__process_cgroup,
 481                 .ordered_events = true,
 482                 .ordering_requires_timestamps = true,
 483         },
 484 };
 485
 486 static struct evsel *evsel_match(struct evsel *evsel,
 487                                       struct evlist *evlist)
 488 {
 489         struct evsel *e;
 490
 491         evlist__for_each_entry(evlist, e) {
 492                 if (evsel__match2(evsel, e))
 493                         return e;
 494         }
 495
 496         return NULL;
 497 }
 498
 499 static void evlist__collapse_resort(struct evlist *evlist)
 500 {
 501         struct evsel *evsel;
 502
 503         evlist__for_each_entry(evlist, evsel) {
 504                 struct hists *hists = evsel__hists(evsel);
 505
 506                 hists__collapse_resort(hists, NULL);
 507         }
 508 }
 509
 510 static struct data__file *fmt_to_data_file(struct perf_hpp_fmt *fmt)
 511 {
 512         struct diff_hpp_fmt *dfmt = container_of(fmt, struct diff_hpp_fmt, fmt);
 513         void *ptr = dfmt - dfmt->idx;
 514         struct data__file *d = container_of(ptr, struct data__file, fmt);
 515
 516         return d;
 517 }
 518
 519 static struct hist_entry*
 520 get_pair_data(struct hist_entry *he, struct data__file *d)
 521 {
 522         if (hist_entry__has_pairs(he)) {
 523                 struct hist_entry *pair;
 524
 525                 list_for_each_entry(pair, &he->pairs.head, pairs.node)
 526                         if (pair->hists == d->hists)
 527                                 return pair;
 528         }
 529
 530         return NULL;
 531 }
 532
 533 static struct hist_entry*
 534 get_pair_fmt(struct hist_entry *he, struct diff_hpp_fmt *dfmt)
 535 {
 536         struct data__file *d = fmt_to_data_file(&dfmt->fmt);
 537
 538         return get_pair_data(he, d);
 539 }
 540
 541 static void hists__baseline_only(struct hists *hists)
 542 {
 543         struct rb_root_cached *root;
 544         struct rb_node *next;
 545
 546         if (hists__has(hists, need_collapse))
 547                 root = &hists->entries_collapsed;
 548         else
 549                 root = hists->entries_in;
 550
 551         next = rb_first_cached(root);
 552         while (next != NULL) {
 553                 struct hist_entry *he = rb_entry(next, struct hist_entry, rb_node_in);
 554
 555                 next = rb_next(&he->rb_node_in);
 556                 if (!hist_entry__next_pair(he)) {
 557                         rb_erase_cached(&he->rb_node_in, root);
 558                         hist_entry__delete(he);
 559                 }
 560         }
 561 }
 562
 563 static int64_t block_cycles_diff_cmp(struct hist_entry *left,
 564                                      struct hist_entry *right)
 565 {
 566         bool pairs_left  = hist_entry__has_pairs(left);
 567         bool pairs_right = hist_entry__has_pairs(right);
 568         s64 l, r;
 569
 570         if (!pairs_left && !pairs_right)
 571                 return 0;
 572
 573         l = llabs(left->diff.cycles);
 574         r = llabs(right->diff.cycles);
 575         return r - l;
 576 }
 577
 578 static int64_t block_sort(struct perf_hpp_fmt *fmt __maybe_unused,
 579                           struct hist_entry *left, struct hist_entry *right)
 580 {
 581         return block_cycles_diff_cmp(right, left);
 582 }
 583
 584 static void init_block_hist(struct block_hist *bh)
 585 {
 586         __hists__init(&bh->block_hists, &bh->block_list);
 587         perf_hpp_list__init(&bh->block_list);
 588
 589         INIT_LIST_HEAD(&bh->block_fmt.list);
 590         INIT_LIST_HEAD(&bh->block_fmt.sort_list);
 591         bh->block_fmt.cmp = block_info__cmp;
 592         bh->block_fmt.sort = block_sort;
 593         perf_hpp_list__register_sort_field(&bh->block_list,
 594                                            &bh->block_fmt);
 595         bh->valid = true;
 596 }
 597
 598 static struct hist_entry *get_block_pair(struct hist_entry *he,
 599                                          struct hists *hists_pair)
 600 {
 601         struct rb_root_cached *root = hists_pair->entries_in;
 602         struct rb_node *next = rb_first_cached(root);
 603         int64_t cmp;
 604
 605         while (next != NULL) {
 606                 struct hist_entry *he_pair = rb_entry(next, struct hist_entry,
 607                                                       rb_node_in);
 608
 609                 next = rb_next(&he_pair->rb_node_in);
 610
 611                 cmp = __block_info__cmp(he_pair, he);
 612                 if (!cmp)
 613                         return he_pair;
 614         }
 615
 616         return NULL;
 617 }
 618
 619 static void init_spark_values(unsigned long *svals, int num)
 620 {
 621         for (int i = 0; i < num; i++)
 622                 svals[i] = 0;
 623 }
 624
 625 static void update_spark_value(unsigned long *svals, int num,
 626                                struct stats *stats, u64 val)
 627 {
 628         int n = stats->n;
 629
 630         if (n < num)
 631                 svals[n] = val;
 632 }
 633
 634 static void compute_cycles_diff(struct hist_entry *he,
 635                                 struct hist_entry *pair)
 636 {
 637         pair->diff.computed = true;
 638         if (pair->block_info->num && he->block_info->num) {
 639                 pair->diff.cycles =
 640                         pair->block_info->cycles_aggr / pair->block_info->num_aggr -
 641                         he->block_info->cycles_aggr / he->block_info->num_aggr;
 642
 643                 if (!cycles_hist)
 644                         return;
 645
 646                 init_stats(&pair->diff.stats);
 647                 init_spark_values(pair->diff.svals, NUM_SPARKS);
 648
 649                 for (int i = 0; i < pair->block_info->num; i++) {
 650                         u64 val;
 651
 652                         if (i >= he->block_info->num || i >= NUM_SPARKS)
 653                                 break;
 654
 655                         val = llabs(pair->block_info->cycles_spark[i] -
 656                                      he->block_info->cycles_spark[i]);
 657
 658                         update_spark_value(pair->diff.svals, NUM_SPARKS,
 659                                            &pair->diff.stats, val);
 660                         update_stats(&pair->diff.stats, val);
 661                 }
 662         }
 663 }
 664
 665 static void block_hists_match(struct hists *hists_base,
 666                               struct hists *hists_pair)
 667 {
 668         struct rb_root_cached *root = hists_base->entries_in;
 669         struct rb_node *next = rb_first_cached(root);
 670
 671         while (next != NULL) {
 672                 struct hist_entry *he = rb_entry(next, struct hist_entry,
 673                                                  rb_node_in);
 674                 struct hist_entry *pair = get_block_pair(he, hists_pair);
 675
 676                 next = rb_next(&he->rb_node_in);
 677
 678                 if (pair) {
 679                         hist_entry__add_pair(pair, he);
 680                         compute_cycles_diff(he, pair);
 681                 }
 682         }
 683 }
 684
 685 static void hists__precompute(struct hists *hists)
 686 {
 687         struct rb_root_cached *root;
 688         struct rb_node *next;
 689
 690         if (hists__has(hists, need_collapse))
 691                 root = &hists->entries_collapsed;
 692         else
 693                 root = hists->entries_in;
 694
 695         next = rb_first_cached(root);
 696         while (next != NULL) {
 697                 struct block_hist *bh, *pair_bh;
 698                 struct hist_entry *he, *pair;
 699                 struct data__file *d;
 700                 int i;
 701
 702                 he   = rb_entry(next, struct hist_entry, rb_node_in);
 703                 next = rb_next(&he->rb_node_in);
 704
 705                 if (compute == COMPUTE_CYCLES) {
 706                         bh = container_of(he, struct block_hist, he);
 707                         init_block_hist(bh);
 708                         block_info__process_sym(he, bh, NULL, 0);
 709                 }
 710
 711                 data__for_each_file_new(i, d) {
 712                         pair = get_pair_data(he, d);
 713                         if (!pair)
 714                                 continue;
 715
 716                         switch (compute) {
 717                         case COMPUTE_DELTA:
 718                         case COMPUTE_DELTA_ABS:
 719                                 compute_delta(he, pair);
 720                                 break;
 721                         case COMPUTE_RATIO:
 722                                 compute_ratio(he, pair);
 723                                 break;
 724                         case COMPUTE_WEIGHTED_DIFF:
 725                                 compute_wdiff(he, pair);
 726                                 break;
 727                         case COMPUTE_CYCLES:
 728                                 pair_bh = container_of(pair, struct block_hist,
 729                                                        he);
 730                                 init_block_hist(pair_bh);
 731                                 block_info__process_sym(pair, pair_bh, NULL, 0);
 732
 733                                 bh = container_of(he, struct block_hist, he);
 734
 735                                 if (bh->valid && pair_bh->valid) {
 736                                         block_hists_match(&bh->block_hists,
 737                                                           &pair_bh->block_hists);
 738                                         hists__output_resort(&pair_bh->block_hists,
 739                                                              NULL);
 740                                 }
 741                                 break;
 742                         default:
 743                                 BUG_ON(1);
 744                         }
 745                 }
 746         }
 747 }
 748
 749 static int64_t cmp_doubles(double l, double r)
 750 {
 751         if (l > r)
 752                 return -1;
 753         else if (l < r)
 754                 return 1;
 755         else
 756                 return 0;
 757 }
 758
 759 static int64_t
 760 __hist_entry__cmp_compute(struct hist_entry *left, struct hist_entry *right,
 761                         int c)
 762 {
 763         switch (c) {
 764         case COMPUTE_DELTA:
 765         {
 766                 double l = left->diff.period_ratio_delta;
 767                 double r = right->diff.period_ratio_delta;
 768
 769                 return cmp_doubles(l, r);
 770         }
 771         case COMPUTE_DELTA_ABS:
 772         {
 773                 double l = fabs(left->diff.period_ratio_delta);
 774                 double r = fabs(right->diff.period_ratio_delta);
 775
 776                 return cmp_doubles(l, r);
 777         }
 778         case COMPUTE_RATIO:
 779         {
 780                 double l = left->diff.period_ratio;
 781                 double r = right->diff.period_ratio;
 782
 783                 return cmp_doubles(l, r);
 784         }
 785         case COMPUTE_WEIGHTED_DIFF:
 786         {
 787                 s64 l = left->diff.wdiff;
 788                 s64 r = right->diff.wdiff;
 789
 790                 return r - l;
 791         }
 792         default:
 793                 BUG_ON(1);
 794         }
 795
 796         return 0;
 797 }
 798
 799 static int64_t
 800 hist_entry__cmp_compute(struct hist_entry *left, struct hist_entry *right,
 801                         int c, int sort_idx)
 802 {
 803         bool pairs_left  = hist_entry__has_pairs(left);
 804         bool pairs_right = hist_entry__has_pairs(right);
 805         struct hist_entry *p_right, *p_left;
 806
 807         if (!pairs_left && !pairs_right)
 808                 return 0;
 809
 810         if (!pairs_left || !pairs_right)
 811                 return pairs_left ? -1 : 1;
 812
 813         p_left  = get_pair_data(left,  &data__files[sort_idx]);
 814         p_right = get_pair_data(right, &data__files[sort_idx]);
 815
 816         if (!p_left && !p_right)
 817                 return 0;
 818
 819         if (!p_left || !p_right)
 820                 return p_left ? -1 : 1;
 821
 822         /*
 823          * We have 2 entries of same kind, let's
 824          * make the data comparison.
 825          */
 826         return __hist_entry__cmp_compute(p_left, p_right, c);
 827 }
 828
 829 static int64_t
 830 hist_entry__cmp_compute_idx(struct hist_entry *left, struct hist_entry *right,
 831                             int c, int sort_idx)
 832 {
 833         struct hist_entry *p_right, *p_left;
 834
 835         p_left  = get_pair_data(left,  &data__files[sort_idx]);
 836         p_right = get_pair_data(right, &data__files[sort_idx]);
 837
 838         if (!p_left && !p_right)
 839                 return 0;
 840
 841         if (!p_left || !p_right)
 842                 return p_left ? -1 : 1;
 843
 844         if (c != COMPUTE_DELTA && c != COMPUTE_DELTA_ABS) {
 845                 /*
 846                  * The delta can be computed without the baseline, but
 847                  * others are not.  Put those entries which have no
 848                  * values below.
 849                  */
 850                 if (left->dummy && right->dummy)
 851                         return 0;
 852
 853                 if (left->dummy || right->dummy)
 854                         return left->dummy ? 1 : -1;
 855         }
 856
 857         return __hist_entry__cmp_compute(p_left, p_right, c);
 858 }
 859
 860 static int64_t
 861 hist_entry__cmp_nop(struct perf_hpp_fmt *fmt __maybe_unused,
 862                     struct hist_entry *left __maybe_unused,
 863                     struct hist_entry *right __maybe_unused)
 864 {
 865         return 0;
 866 }
 867
 868 static int64_t
 869 hist_entry__cmp_baseline(struct perf_hpp_fmt *fmt __maybe_unused,
 870                          struct hist_entry *left, struct hist_entry *right)
 871 {
 872         if (left->stat.period == right->stat.period)
 873                 return 0;
 874         return left->stat.period > right->stat.period ? 1 : -1;
 875 }
 876
 877 static int64_t
 878 hist_entry__cmp_delta(struct perf_hpp_fmt *fmt,
 879                       struct hist_entry *left, struct hist_entry *right)
 880 {
 881         struct data__file *d = fmt_to_data_file(fmt);
 882
 883         return hist_entry__cmp_compute(right, left, COMPUTE_DELTA, d->idx);
 884 }
 885
 886 static int64_t
 887 hist_entry__cmp_delta_abs(struct perf_hpp_fmt *fmt,
 888                       struct hist_entry *left, struct hist_entry *right)
 889 {
 890         struct data__file *d = fmt_to_data_file(fmt);
 891
 892         return hist_entry__cmp_compute(right, left, COMPUTE_DELTA_ABS, d->idx);
 893 }
 894
 895 static int64_t
 896 hist_entry__cmp_ratio(struct perf_hpp_fmt *fmt,
 897                       struct hist_entry *left, struct hist_entry *right)
 898 {
 899         struct data__file *d = fmt_to_data_file(fmt);
 900
 901         return hist_entry__cmp_compute(right, left, COMPUTE_RATIO, d->idx);
 902 }
 903
 904 static int64_t
 905 hist_entry__cmp_wdiff(struct perf_hpp_fmt *fmt,
 906                       struct hist_entry *left, struct hist_entry *right)
 907 {
 908         struct data__file *d = fmt_to_data_file(fmt);
 909
 910         return hist_entry__cmp_compute(right, left, COMPUTE_WEIGHTED_DIFF, d->idx);
 911 }
 912
 913 static int64_t
 914 hist_entry__cmp_delta_idx(struct perf_hpp_fmt *fmt __maybe_unused,
 915                           struct hist_entry *left, struct hist_entry *right)
 916 {
 917         return hist_entry__cmp_compute_idx(right, left, COMPUTE_DELTA,
 918                                            sort_compute);
 919 }
 920
 921 static int64_t
 922 hist_entry__cmp_delta_abs_idx(struct perf_hpp_fmt *fmt __maybe_unused,
 923                               struct hist_entry *left, struct hist_entry *right)
 924 {
 925         return hist_entry__cmp_compute_idx(right, left, COMPUTE_DELTA_ABS,
 926                                            sort_compute);
 927 }
 928
 929 static int64_t
 930 hist_entry__cmp_ratio_idx(struct perf_hpp_fmt *fmt __maybe_unused,
 931                           struct hist_entry *left, struct hist_entry *right)
 932 {
 933         return hist_entry__cmp_compute_idx(right, left, COMPUTE_RATIO,
 934                                            sort_compute);
 935 }
 936
 937 static int64_t
 938 hist_entry__cmp_wdiff_idx(struct perf_hpp_fmt *fmt __maybe_unused,
 939                           struct hist_entry *left, struct hist_entry *right)
 940 {
 941         return hist_entry__cmp_compute_idx(right, left, COMPUTE_WEIGHTED_DIFF,
 942                                            sort_compute);
 943 }
 944
 945 static void hists__process(struct hists *hists)
 946 {
 947         if (show_baseline_only)
 948                 hists__baseline_only(hists);
 949
 950         hists__precompute(hists);
 951         hists__output_resort(hists, NULL);
 952
 953         if (compute == COMPUTE_CYCLES)
 954                 symbol_conf.report_block = true;
 955
 956         hists__fprintf(hists, !quiet, 0, 0, 0, stdout,
 957                        !symbol_conf.use_callchain);
 958 }
 959
 960 static void data__fprintf(void)
 961 {
 962         struct data__file *d;
 963         int i;
 964
 965         fprintf(stdout, "# Data files:\n");
 966
 967         data__for_each_file(i, d)
 968                 fprintf(stdout, "#  [%d] %s %s\n",
 969                         d->idx, d->data.path,
 970                         !d->idx ? "(Baseline)" : "");
 971
 972         fprintf(stdout, "#\n");
 973 }
 974
 975 static void data_process(void)
 976 {
 977         struct evlist *evlist_base = data__files[0].session->evlist;
 978         struct evsel *evsel_base;
 979         bool first = true;
 980
 981         evlist__for_each_entry(evlist_base, evsel_base) {
 982                 struct hists *hists_base = evsel__hists(evsel_base);
 983                 struct data__file *d;
 984                 int i;
 985
 986                 data__for_each_file_new(i, d) {
 987                         struct evlist *evlist = d->session->evlist;
 988                         struct evsel *evsel;
 989                         struct hists *hists;
 990
 991                         evsel = evsel_match(evsel_base, evlist);
 992                         if (!evsel)
 993                                 continue;
 994
 995                         hists = evsel__hists(evsel);
 996                         d->hists = hists;
 997
 998                         hists__match(hists_base, hists);
 999
1000                         if (!show_baseline_only)
1001                                 hists__link(hists_base, hists);
1002                 }
1003
1004                 if (!quiet) {
1005                         fprintf(stdout, "%s# Event '%s'\n#\n", first ? "" : "\n",
1006                                 evsel__name(evsel_base));
1007                 }
1008
1009                 first = false;
1010
1011                 if (verbose > 0 || ((data__files_cnt > 2) && !quiet))
1012                         data__fprintf();
1013
1014                 /* Don't sort callchain for perf diff */
1015                 evsel__reset_sample_bit(evsel_base, CALLCHAIN);
1016
1017                 hists__process(hists_base);
1018         }
1019 }
1020
1021 static int process_base_stream(struct data__file *data_base,
1022                                struct data__file *data_pair,
1023                                const char *title __maybe_unused)
1024 {
1025         struct evlist *evlist_base = data_base->session->evlist;
1026         struct evlist *evlist_pair = data_pair->session->evlist;
1027         struct evsel *evsel_base, *evsel_pair;
1028         struct evsel_streams *es_base, *es_pair;
1029
1030         evlist__for_each_entry(evlist_base, evsel_base) {
1031                 evsel_pair = evsel_match(evsel_base, evlist_pair);
1032                 if (!evsel_pair)
1033                         continue;
1034
1035                 es_base = evsel_streams__entry(data_base->evlist_streams,
1036                                                evsel_base->core.idx);
1037                 if (!es_base)
1038                         return -1;
1039
1040                 es_pair = evsel_streams__entry(data_pair->evlist_streams,
1041                                                evsel_pair->core.idx);
1042                 if (!es_pair)
1043                         return -1;
1044
1045                 evsel_streams__match(es_base, es_pair);
1046                 evsel_streams__report(es_base, es_pair);
1047         }
1048
1049         return 0;
1050 }
1051
1052 static void stream_process(void)
1053 {
1054         /*
1055          * Stream comparison only supports two data files.
1056          * perf.data.old and perf.data. data__files[0] is perf.data.old,
1057          * data__files[1] is perf.data.
1058          */
1059         process_base_stream(&data__files[0], &data__files[1],
1060                             "# Output based on old perf data:\n#\n");
1061 }
1062
1063 static void data__free(struct data__file *d)
1064 {
1065         int col;
1066
1067         if (d->evlist_streams)
1068                 evlist_streams__delete(d->evlist_streams);
1069
1070         for (col = 0; col < PERF_HPP_DIFF__MAX_INDEX; col++) {
1071                 struct diff_hpp_fmt *fmt = &d->fmt[col];
1072
1073                 zfree(&fmt->header);
1074         }
1075 }
1076
1077 static int abstime_str_dup(char **pstr)
1078 {
1079         char *str = NULL;
1080
1081         if (pdiff.time_str && strchr(pdiff.time_str, ':')) {
1082                 str = strdup(pdiff.time_str);
1083                 if (!str)
1084                         return -ENOMEM;
1085         }
1086
1087         *pstr = str;
1088         return 0;
1089 }
1090
1091 static int parse_absolute_time(struct data__file *d, char **pstr)
1092 {
1093         char *p = *pstr;
1094         int ret;
1095
1096         /*
1097          * Absolute timestamp for one file has the format: a.b,c.d
1098          * For multiple files, the format is: a.b,c.d:a.b,c.d
1099          */
1100         p = strchr(*pstr, ':');
1101         if (p) {
1102                 if (p == *pstr) {
1103                         pr_err("Invalid time string\n");
1104                         return -EINVAL;
1105                 }
1106
1107                 *p = 0;
1108                 p++;
1109                 if (*p == 0) {
1110                         pr_err("Invalid time string\n");
1111                         return -EINVAL;
1112                 }
1113         }
1114
1115         ret = perf_time__parse_for_ranges(*pstr, d->session,
1116                                           &pdiff.ptime_range,
1117                                           &pdiff.range_size,
1118                                           &pdiff.range_num);
1119         if (ret < 0)
1120                 return ret;
1121
1122         if (!p || *p == 0)
1123                 *pstr = NULL;
1124         else
1125                 *pstr = p;
1126
1127         return ret;
1128 }
1129
1130 static int parse_percent_time(struct data__file *d)
1131 {
1132         int ret;
1133
1134         ret = perf_time__parse_for_ranges(pdiff.time_str, d->session,
1135                                           &pdiff.ptime_range,
1136                                           &pdiff.range_size,
1137                                           &pdiff.range_num);
1138         return ret;
1139 }
1140
1141 static int parse_time_str(struct data__file *d, char *abstime_ostr,
1142                            char **pabstime_tmp)
1143 {
1144         int ret = 0;
1145
1146         if (abstime_ostr)
1147                 ret = parse_absolute_time(d, pabstime_tmp);
1148         else if (pdiff.time_str)
1149                 ret = parse_percent_time(d);
1150
1151         return ret;
1152 }
1153
1154 static int check_file_brstack(void)
1155 {
1156         struct data__file *d;
1157         bool has_br_stack;
1158         int i;
1159
1160         data__for_each_file(i, d) {
1161                 d->session = perf_session__new(&d->data, &pdiff.tool);
1162                 if (IS_ERR(d->session)) {
1163                         pr_err("Failed to open %s\n", d->data.path);
1164                         return PTR_ERR(d->session);
1165                 }
1166
1167                 has_br_stack = perf_header__has_feat(&d->session->header,
1168                                                      HEADER_BRANCH_STACK);
1169                 perf_session__delete(d->session);
1170                 if (!has_br_stack)
1171                         return 0;
1172         }
1173
1174         /* Set only all files having branch stacks */
1175         pdiff.has_br_stack = true;
1176         return 0;
1177 }
1178
1179 static int __cmd_diff(void)
1180 {
1181         struct data__file *d;
1182         int ret, i;
1183         char *abstime_ostr, *abstime_tmp;
1184
1185         ret = abstime_str_dup(&abstime_ostr);
1186         if (ret)
1187                 return ret;
1188
1189         abstime_tmp = abstime_ostr;
1190         ret = -EINVAL;
1191
1192         data__for_each_file(i, d) {
1193                 d->session = perf_session__new(&d->data, &pdiff.tool);
1194                 if (IS_ERR(d->session)) {
1195                         ret = PTR_ERR(d->session);
1196                         pr_err("Failed to open %s\n", d->data.path);
1197                         goto out_delete;
1198                 }
1199
1200                 if (pdiff.time_str) {
1201                         ret = parse_time_str(d, abstime_ostr, &abstime_tmp);
1202                         if (ret < 0)
1203                                 goto out_delete;
1204                 }
1205
1206                 if (cpu_list) {
1207                         ret = perf_session__cpu_bitmap(d->session, cpu_list,
1208                                                        cpu_bitmap);
1209                         if (ret < 0)
1210                                 goto out_delete;
1211                 }
1212
1213                 ret = perf_session__process_events(d->session);
1214                 if (ret) {
1215                         pr_err("Failed to process %s\n", d->data.path);
1216                         goto out_delete;
1217                 }
1218
1219                 evlist__collapse_resort(d->session->evlist);
1220
1221                 if (pdiff.ptime_range)
1222                         zfree(&pdiff.ptime_range);
1223
1224                 if (compute == COMPUTE_STREAM) {
1225                         d->evlist_streams = evlist__create_streams(
1226                                                 d->session->evlist, 5);
1227                         if (!d->evlist_streams) {
1228                                 ret = -ENOMEM;
1229                                 goto out_delete;
1230                         }
1231                 }
1232         }
1233
1234         if (compute == COMPUTE_STREAM)
1235                 stream_process();
1236         else
1237                 data_process();
1238
1239  out_delete:
1240         data__for_each_file(i, d) {
1241                 if (!IS_ERR(d->session))
1242                         perf_session__delete(d->session);
1243                 data__free(d);
1244         }
1245
1246         free(data__files);
1247
1248         if (pdiff.ptime_range)
1249                 zfree(&pdiff.ptime_range);
1250
1251         if (abstime_ostr)
1252                 free(abstime_ostr);
1253
1254         return ret;
1255 }
1256
1257 static const char * const diff_usage[] = {
1258         "perf diff [<options>] [old_file] [new_file]",
1259         NULL,
1260 };
1261
1262 static const struct option options[] = {
1263         OPT_INCR('v', "verbose", &verbose,
1264                     "be more verbose (show symbol address, etc)"),
1265         OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any warnings or messages"),
1266         OPT_BOOLEAN('b', "baseline-only", &show_baseline_only,
1267                     "Show only items with match in baseline"),
1268         OPT_CALLBACK('c', "compute", &compute,
1269                      "delta,delta-abs,ratio,wdiff:w1,w2 (default delta-abs),cycles",
1270                      "Entries differential computation selection",
1271                      setup_compute),
1272         OPT_BOOLEAN('p', "period", &show_period,
1273                     "Show period values."),
1274         OPT_BOOLEAN('F', "formula", &show_formula,
1275                     "Show formula."),
1276         OPT_BOOLEAN(0, "cycles-hist", &cycles_hist,
1277                     "Show cycles histogram and standard deviation "
1278                     "- WARNING: use only with -c cycles."),
1279         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1280                     "dump raw trace in ASCII"),
1281         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
1282         OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1283                    "file", "kallsyms pathname"),
1284         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
1285                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
1286         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1287                    "only consider symbols in these dsos"),
1288         OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1289                    "only consider symbols in these comms"),
1290         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1291                    "only consider these symbols"),
1292         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1293                    "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1294                    " Please refer the man page for the complete list."),
1295         OPT_STRING_NOEMPTY('t', "field-separator", &symbol_conf.field_sep, "separator",
1296                    "separator for columns, no spaces will be added between "
1297                    "columns '.' is reserved."),
1298         OPT_CALLBACK(0, "symfs", NULL, "directory",
1299                      "Look for files with symbols relative to this directory",
1300                      symbol__config_symfs),
1301         OPT_UINTEGER('o', "order", &sort_compute, "Specify compute sorting."),
1302         OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1303                      "How to display percentage of filtered entries", parse_filter_percentage),
1304         OPT_STRING(0, "time", &pdiff.time_str, "str",
1305                    "Time span (time percent or absolute timestamp)"),
1306         OPT_STRING(0, "cpu", &cpu_list, "cpu", "list of cpus to profile"),
1307         OPT_STRING(0, "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
1308                    "only consider symbols in these pids"),
1309         OPT_STRING(0, "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
1310                    "only consider symbols in these tids"),
1311         OPT_BOOLEAN(0, "stream", &pdiff.stream,
1312                     "Enable hot streams comparison."),
1313         OPT_END()
1314 };
1315
1316 static double baseline_percent(struct hist_entry *he)
1317 {
1318         u64 total = hists__total_period(he->hists);
1319
1320         return 100.0 * he->stat.period / total;
1321 }
1322
1323 static int hpp__color_baseline(struct perf_hpp_fmt *fmt,
1324                                struct perf_hpp *hpp, struct hist_entry *he)
1325 {
1326         struct diff_hpp_fmt *dfmt =
1327                 container_of(fmt, struct diff_hpp_fmt, fmt);
1328         double percent = baseline_percent(he);
1329         char pfmt[20] = " ";
1330
1331         if (!he->dummy) {
1332                 scnprintf(pfmt, 20, "%%%d.2f%%%%", dfmt->header_width - 1);
1333                 return percent_color_snprintf(hpp->buf, hpp->size,
1334                                               pfmt, percent);
1335         } else
1336                 return scnprintf(hpp->buf, hpp->size, "%*s",
1337                                  dfmt->header_width, pfmt);
1338 }
1339
1340 static int hpp__entry_baseline(struct hist_entry *he, char *buf, size_t size)
1341 {
1342         double percent = baseline_percent(he);
1343         const char *fmt = symbol_conf.field_sep ? "%.2f" : "%6.2f%%";
1344         int ret = 0;
1345
1346         if (!he->dummy)
1347                 ret = scnprintf(buf, size, fmt, percent);
1348
1349         return ret;
1350 }
1351
1352 static int cycles_printf(struct hist_entry *he, struct hist_entry *pair,
1353                          struct perf_hpp *hpp, int width)
1354 {
1355         struct block_hist *bh = container_of(he, struct block_hist, he);
1356         struct block_hist *bh_pair = container_of(pair, struct block_hist, he);
1357         struct hist_entry *block_he;
1358         struct block_info *bi;
1359         char buf[128];
1360         char *start_line, *end_line;
1361
1362         block_he = hists__get_entry(&bh_pair->block_hists, bh->block_idx);
1363         if (!block_he) {
1364                 hpp->skip = true;
1365                 return 0;
1366         }
1367
1368         /*
1369          * Avoid printing the warning "addr2line_init failed for ..."
1370          */
1371         symbol_conf.disable_add2line_warn = true;
1372
1373         bi = block_he->block_info;
1374
1375         start_line = map__srcline(he->ms.map, bi->sym->start + bi->start,
1376                                   he->ms.sym);
1377
1378         end_line = map__srcline(he->ms.map, bi->sym->start + bi->end,
1379                                 he->ms.sym);
1380
1381         if (start_line != SRCLINE_UNKNOWN &&
1382             end_line != SRCLINE_UNKNOWN) {
1383                 scnprintf(buf, sizeof(buf), "[%s -> %s] %4ld",
1384                           start_line, end_line, block_he->diff.cycles);
1385         } else {
1386                 scnprintf(buf, sizeof(buf), "[%7lx -> %7lx] %4ld",
1387                           bi->start, bi->end, block_he->diff.cycles);
1388         }
1389
1390         zfree_srcline(&start_line);
1391         zfree_srcline(&end_line);
1392
1393         return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1394 }
1395
1396 static int __hpp__color_compare(struct perf_hpp_fmt *fmt,
1397                                 struct perf_hpp *hpp, struct hist_entry *he,
1398                                 int comparison_method)
1399 {
1400         struct diff_hpp_fmt *dfmt =
1401                 container_of(fmt, struct diff_hpp_fmt, fmt);
1402         struct hist_entry *pair = get_pair_fmt(he, dfmt);
1403         double diff;
1404         s64 wdiff;
1405         char pfmt[20] = " ";
1406
1407         if (!pair) {
1408                 if (comparison_method == COMPUTE_CYCLES) {
1409                         struct block_hist *bh;
1410
1411                         bh = container_of(he, struct block_hist, he);
1412                         if (bh->block_idx)
1413                                 hpp->skip = true;
1414                 }
1415
1416                 goto no_print;
1417         }
1418
1419         switch (comparison_method) {
1420         case COMPUTE_DELTA:
1421                 if (pair->diff.computed)
1422                         diff = pair->diff.period_ratio_delta;
1423                 else
1424                         diff = compute_delta(he, pair);
1425
1426                 scnprintf(pfmt, 20, "%%%+d.2f%%%%", dfmt->header_width - 1);
1427                 return percent_color_snprintf(hpp->buf, hpp->size,
1428                                         pfmt, diff);
1429         case COMPUTE_RATIO:
1430                 if (he->dummy)
1431                         goto dummy_print;
1432                 if (pair->diff.computed)
1433                         diff = pair->diff.period_ratio;
1434                 else
1435                         diff = compute_ratio(he, pair);
1436
1437                 scnprintf(pfmt, 20, "%%%d.6f", dfmt->header_width);
1438                 return value_color_snprintf(hpp->buf, hpp->size,
1439                                         pfmt, diff);
1440         case COMPUTE_WEIGHTED_DIFF:
1441                 if (he->dummy)
1442                         goto dummy_print;
1443                 if (pair->diff.computed)
1444                         wdiff = pair->diff.wdiff;
1445                 else
1446                         wdiff = compute_wdiff(he, pair);
1447
1448                 scnprintf(pfmt, 20, "%%14ld", dfmt->header_width);
1449                 return color_snprintf(hpp->buf, hpp->size,
1450                                 get_percent_color(wdiff),
1451                                 pfmt, wdiff);
1452         case COMPUTE_CYCLES:
1453                 return cycles_printf(he, pair, hpp, dfmt->header_width);
1454         default:
1455                 BUG_ON(1);
1456         }
1457 dummy_print:
1458         return scnprintf(hpp->buf, hpp->size, "%*s",
1459                         dfmt->header_width, "N/A");
1460 no_print:
1461         return scnprintf(hpp->buf, hpp->size, "%*s",
1462                         dfmt->header_width, pfmt);
1463 }
1464
1465 static int hpp__color_delta(struct perf_hpp_fmt *fmt,
1466                         struct perf_hpp *hpp, struct hist_entry *he)
1467 {
1468         return __hpp__color_compare(fmt, hpp, he, COMPUTE_DELTA);
1469 }
1470
1471 static int hpp__color_ratio(struct perf_hpp_fmt *fmt,
1472                         struct perf_hpp *hpp, struct hist_entry *he)
1473 {
1474         return __hpp__color_compare(fmt, hpp, he, COMPUTE_RATIO);
1475 }
1476
1477 static int hpp__color_wdiff(struct perf_hpp_fmt *fmt,
1478                         struct perf_hpp *hpp, struct hist_entry *he)
1479 {
1480         return __hpp__color_compare(fmt, hpp, he, COMPUTE_WEIGHTED_DIFF);
1481 }
1482
1483 static int hpp__color_cycles(struct perf_hpp_fmt *fmt,
1484                              struct perf_hpp *hpp, struct hist_entry *he)
1485 {
1486         return __hpp__color_compare(fmt, hpp, he, COMPUTE_CYCLES);
1487 }
1488
1489 static int all_zero(unsigned long *vals, int len)
1490 {
1491         int i;
1492
1493         for (i = 0; i < len; i++)
1494                 if (vals[i] != 0)
1495                         return 0;
1496         return 1;
1497 }
1498
1499 static int print_cycles_spark(char *bf, int size, unsigned long *svals, u64 n)
1500 {
1501         int printed;
1502
1503         if (n <= 1)
1504                 return 0;
1505
1506         if (n > NUM_SPARKS)
1507                 n = NUM_SPARKS;
1508         if (all_zero(svals, n))
1509                 return 0;
1510
1511         printed = print_spark(bf, size, svals, n);
1512         printed += scnprintf(bf + printed, size - printed, " ");
1513         return printed;
1514 }
1515
1516 static int hpp__color_cycles_hist(struct perf_hpp_fmt *fmt,
1517                             struct perf_hpp *hpp, struct hist_entry *he)
1518 {
1519         struct diff_hpp_fmt *dfmt =
1520                 container_of(fmt, struct diff_hpp_fmt, fmt);
1521         struct hist_entry *pair = get_pair_fmt(he, dfmt);
1522         struct block_hist *bh = container_of(he, struct block_hist, he);
1523         struct block_hist *bh_pair;
1524         struct hist_entry *block_he;
1525         char spark[32], buf[128];
1526         double r;
1527         int ret, pad;
1528
1529         if (!pair) {
1530                 if (bh->block_idx)
1531                         hpp->skip = true;
1532
1533                 goto no_print;
1534         }
1535
1536         bh_pair = container_of(pair, struct block_hist, he);
1537
1538         block_he = hists__get_entry(&bh_pair->block_hists, bh->block_idx);
1539         if (!block_he) {
1540                 hpp->skip = true;
1541                 goto no_print;
1542         }
1543
1544         ret = print_cycles_spark(spark, sizeof(spark), block_he->diff.svals,
1545                                  block_he->diff.stats.n);
1546
1547         r = rel_stddev_stats(stddev_stats(&block_he->diff.stats),
1548                              avg_stats(&block_he->diff.stats));
1549
1550         if (ret) {
1551                 /*
1552                  * Padding spaces if number of sparks less than NUM_SPARKS
1553                  * otherwise the output is not aligned.
1554                  */
1555                 pad = NUM_SPARKS - ((ret - 1) / 3);
1556                 scnprintf(buf, sizeof(buf), "%s%5.1f%% %s", "\u00B1", r, spark);
1557                 ret = scnprintf(hpp->buf, hpp->size, "%*s",
1558                                 dfmt->header_width, buf);
1559
1560                 if (pad) {
1561                         ret += scnprintf(hpp->buf + ret, hpp->size - ret,
1562                                          "%-*s", pad, " ");
1563                 }
1564
1565                 return ret;
1566         }
1567
1568 no_print:
1569         return scnprintf(hpp->buf, hpp->size, "%*s",
1570                         dfmt->header_width, " ");
1571 }
1572
1573 static void
1574 hpp__entry_unpair(struct hist_entry *he, int idx, char *buf, size_t size)
1575 {
1576         switch (idx) {
1577         case PERF_HPP_DIFF__PERIOD_BASELINE:
1578                 scnprintf(buf, size, "%" PRIu64, he->stat.period);
1579                 break;
1580
1581         default:
1582                 break;
1583         }
1584 }
1585
1586 static void
1587 hpp__entry_pair(struct hist_entry *he, struct hist_entry *pair,
1588                 int idx, char *buf, size_t size)
1589 {
1590         double diff;
1591         double ratio;
1592         s64 wdiff;
1593
1594         switch (idx) {
1595         case PERF_HPP_DIFF__DELTA:
1596         case PERF_HPP_DIFF__DELTA_ABS:
1597                 if (pair->diff.computed)
1598                         diff = pair->diff.period_ratio_delta;
1599                 else
1600                         diff = compute_delta(he, pair);
1601
1602                 scnprintf(buf, size, "%+4.2F%%", diff);
1603                 break;
1604
1605         case PERF_HPP_DIFF__RATIO:
1606                 /* No point for ratio number if we are dummy.. */
1607                 if (he->dummy) {
1608                         scnprintf(buf, size, "N/A");
1609                         break;
1610                 }
1611
1612                 if (pair->diff.computed)
1613                         ratio = pair->diff.period_ratio;
1614                 else
1615                         ratio = compute_ratio(he, pair);
1616
1617                 if (ratio > 0.0)
1618                         scnprintf(buf, size, "%14.6F", ratio);
1619                 break;
1620
1621         case PERF_HPP_DIFF__WEIGHTED_DIFF:
1622                 /* No point for wdiff number if we are dummy.. */
1623                 if (he->dummy) {
1624                         scnprintf(buf, size, "N/A");
1625                         break;
1626                 }
1627
1628                 if (pair->diff.computed)
1629                         wdiff = pair->diff.wdiff;
1630                 else
1631                         wdiff = compute_wdiff(he, pair);
1632
1633                 if (wdiff != 0)
1634                         scnprintf(buf, size, "%14ld", wdiff);
1635                 break;
1636
1637         case PERF_HPP_DIFF__FORMULA:
1638                 formula_fprintf(he, pair, buf, size);
1639                 break;
1640
1641         case PERF_HPP_DIFF__PERIOD:
1642                 scnprintf(buf, size, "%" PRIu64, pair->stat.period);
1643                 break;
1644
1645         default:
1646                 BUG_ON(1);
1647         }
1648 }
1649
1650 static void
1651 __hpp__entry_global(struct hist_entry *he, struct diff_hpp_fmt *dfmt,
1652                     char *buf, size_t size)
1653 {
1654         struct hist_entry *pair = get_pair_fmt(he, dfmt);
1655         int idx = dfmt->idx;
1656
1657         /* baseline is special */
1658         if (idx == PERF_HPP_DIFF__BASELINE)
1659                 hpp__entry_baseline(he, buf, size);
1660         else {
1661                 if (pair)
1662                         hpp__entry_pair(he, pair, idx, buf, size);
1663                 else
1664                         hpp__entry_unpair(he, idx, buf, size);
1665         }
1666 }
1667
1668 static int hpp__entry_global(struct perf_hpp_fmt *_fmt, struct perf_hpp *hpp,
1669                              struct hist_entry *he)
1670 {
1671         struct diff_hpp_fmt *dfmt =
1672                 container_of(_fmt, struct diff_hpp_fmt, fmt);
1673         char buf[MAX_COL_WIDTH] = " ";
1674
1675         __hpp__entry_global(he, dfmt, buf, MAX_COL_WIDTH);
1676
1677         if (symbol_conf.field_sep)
1678                 return scnprintf(hpp->buf, hpp->size, "%s", buf);
1679         else
1680                 return scnprintf(hpp->buf, hpp->size, "%*s",
1681                                  dfmt->header_width, buf);
1682 }
1683
1684 static int hpp__header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1685                        struct hists *hists __maybe_unused,
1686                        int line __maybe_unused,
1687                        int *span __maybe_unused)
1688 {
1689         struct diff_hpp_fmt *dfmt =
1690                 container_of(fmt, struct diff_hpp_fmt, fmt);
1691
1692         BUG_ON(!dfmt->header);
1693         return scnprintf(hpp->buf, hpp->size, dfmt->header);
1694 }
1695
1696 static int hpp__width(struct perf_hpp_fmt *fmt,
1697                       struct perf_hpp *hpp __maybe_unused,
1698                       struct hists *hists __maybe_unused)
1699 {
1700         struct diff_hpp_fmt *dfmt =
1701                 container_of(fmt, struct diff_hpp_fmt, fmt);
1702
1703         BUG_ON(dfmt->header_width <= 0);
1704         return dfmt->header_width;
1705 }
1706
1707 static void init_header(struct data__file *d, struct diff_hpp_fmt *dfmt)
1708 {
1709 #define MAX_HEADER_NAME 100
1710         char buf_indent[MAX_HEADER_NAME];
1711         char buf[MAX_HEADER_NAME];
1712         const char *header = NULL;
1713         int width = 0;
1714
1715         BUG_ON(dfmt->idx >= PERF_HPP_DIFF__MAX_INDEX);
1716         header = columns[dfmt->idx].name;
1717         width  = columns[dfmt->idx].width;
1718
1719         /* Only our defined HPP fmts should appear here. */
1720         BUG_ON(!header);
1721
1722         if (data__files_cnt > 2)
1723                 scnprintf(buf, MAX_HEADER_NAME, "%s/%d", header, d->idx);
1724
1725 #define NAME (data__files_cnt > 2 ? buf : header)
1726         dfmt->header_width = width;
1727         width = (int) strlen(NAME);
1728         if (dfmt->header_width < width)
1729                 dfmt->header_width = width;
1730
1731         scnprintf(buf_indent, MAX_HEADER_NAME, "%*s",
1732                   dfmt->header_width, NAME);
1733
1734         dfmt->header = strdup(buf_indent);
1735 #undef MAX_HEADER_NAME
1736 #undef NAME
1737 }
1738
1739 static void data__hpp_register(struct data__file *d, int idx)
1740 {
1741         struct diff_hpp_fmt *dfmt = &d->fmt[idx];
1742         struct perf_hpp_fmt *fmt = &dfmt->fmt;
1743
1744         dfmt->idx = idx;
1745
1746         fmt->header = hpp__header;
1747         fmt->width  = hpp__width;
1748         fmt->entry  = hpp__entry_global;
1749         fmt->cmp    = hist_entry__cmp_nop;
1750         fmt->collapse = hist_entry__cmp_nop;
1751
1752         /* TODO more colors */
1753         switch (idx) {
1754         case PERF_HPP_DIFF__BASELINE:
1755                 fmt->color = hpp__color_baseline;
1756                 fmt->sort  = hist_entry__cmp_baseline;
1757                 break;
1758         case PERF_HPP_DIFF__DELTA:
1759                 fmt->color = hpp__color_delta;
1760                 fmt->sort  = hist_entry__cmp_delta;
1761                 break;
1762         case PERF_HPP_DIFF__RATIO:
1763                 fmt->color = hpp__color_ratio;
1764                 fmt->sort  = hist_entry__cmp_ratio;
1765                 break;
1766         case PERF_HPP_DIFF__WEIGHTED_DIFF:
1767                 fmt->color = hpp__color_wdiff;
1768                 fmt->sort  = hist_entry__cmp_wdiff;
1769                 break;
1770         case PERF_HPP_DIFF__DELTA_ABS:
1771                 fmt->color = hpp__color_delta;
1772                 fmt->sort  = hist_entry__cmp_delta_abs;
1773                 break;
1774         case PERF_HPP_DIFF__CYCLES:
1775                 fmt->color = hpp__color_cycles;
1776                 fmt->sort  = hist_entry__cmp_nop;
1777                 break;
1778         case PERF_HPP_DIFF__CYCLES_HIST:
1779                 fmt->color = hpp__color_cycles_hist;
1780                 fmt->sort  = hist_entry__cmp_nop;
1781                 break;
1782         default:
1783                 fmt->sort  = hist_entry__cmp_nop;
1784                 break;
1785         }
1786
1787         init_header(d, dfmt);
1788         perf_hpp__column_register(fmt);
1789         perf_hpp__register_sort_field(fmt);
1790 }
1791
1792 static int ui_init(void)
1793 {
1794         struct data__file *d;
1795         struct perf_hpp_fmt *fmt;
1796         int i;
1797
1798         data__for_each_file(i, d) {
1799
1800                 /*
1801                  * Baseline or compute related columns:
1802                  *
1803                  *   PERF_HPP_DIFF__BASELINE
1804                  *   PERF_HPP_DIFF__DELTA
1805                  *   PERF_HPP_DIFF__RATIO
1806                  *   PERF_HPP_DIFF__WEIGHTED_DIFF
1807                  *   PERF_HPP_DIFF__CYCLES
1808                  */
1809                 data__hpp_register(d, i ? compute_2_hpp[compute] :
1810                                           PERF_HPP_DIFF__BASELINE);
1811
1812                 if (cycles_hist && i)
1813                         data__hpp_register(d, PERF_HPP_DIFF__CYCLES_HIST);
1814
1815                 /*
1816                  * And the rest:
1817                  *
1818                  * PERF_HPP_DIFF__FORMULA
1819                  * PERF_HPP_DIFF__PERIOD
1820                  * PERF_HPP_DIFF__PERIOD_BASELINE
1821                  */
1822                 if (show_formula && i)
1823                         data__hpp_register(d, PERF_HPP_DIFF__FORMULA);
1824
1825                 if (show_period)
1826                         data__hpp_register(d, i ? PERF_HPP_DIFF__PERIOD :
1827                                                   PERF_HPP_DIFF__PERIOD_BASELINE);
1828         }
1829
1830         if (!sort_compute)
1831                 return 0;
1832
1833         /*
1834          * Prepend an fmt to sort on columns at 'sort_compute' first.
1835          * This fmt is added only to the sort list but not to the
1836          * output fields list.
1837          *
1838          * Note that this column (data) can be compared twice - one
1839          * for this 'sort_compute' fmt and another for the normal
1840          * diff_hpp_fmt.  But it shouldn't a problem as most entries
1841          * will be sorted out by first try or baseline and comparing
1842          * is not a costly operation.
1843          */
1844         fmt = zalloc(sizeof(*fmt));
1845         if (fmt == NULL) {
1846                 pr_err("Memory allocation failed\n");
1847                 return -1;
1848         }
1849
1850         fmt->cmp      = hist_entry__cmp_nop;
1851         fmt->collapse = hist_entry__cmp_nop;
1852
1853         switch (compute) {
1854         case COMPUTE_DELTA:
1855                 fmt->sort = hist_entry__cmp_delta_idx;
1856                 break;
1857         case COMPUTE_RATIO:
1858                 fmt->sort = hist_entry__cmp_ratio_idx;
1859                 break;
1860         case COMPUTE_WEIGHTED_DIFF:
1861                 fmt->sort = hist_entry__cmp_wdiff_idx;
1862                 break;
1863         case COMPUTE_DELTA_ABS:
1864                 fmt->sort = hist_entry__cmp_delta_abs_idx;
1865                 break;
1866         case COMPUTE_CYCLES:
1867                 /*
1868                  * Should set since 'fmt->sort' is called without
1869                  * checking valid during sorting
1870                  */
1871                 fmt->sort = hist_entry__cmp_nop;
1872                 break;
1873         default:
1874                 BUG_ON(1);
1875         }
1876
1877         perf_hpp__prepend_sort_field(fmt);
1878         return 0;
1879 }
1880
1881 static int data_init(int argc, const char **argv)
1882 {
1883         struct data__file *d;
1884         static const char *defaults[] = {
1885                 "perf.data.old",
1886                 "perf.data",
1887         };
1888         bool use_default = true;
1889         int i;
1890
1891         data__files_cnt = 2;
1892
1893         if (argc) {
1894                 if (argc == 1)
1895                         defaults[1] = argv[0];
1896                 else {
1897                         data__files_cnt = argc;
1898                         use_default = false;
1899                 }
1900         } else if (perf_guest) {
1901                 defaults[0] = "perf.data.host";
1902                 defaults[1] = "perf.data.guest";
1903         }
1904
1905         if (sort_compute >= (unsigned int) data__files_cnt) {
1906                 pr_err("Order option out of limit.\n");
1907                 return -EINVAL;
1908         }
1909
1910         data__files = zalloc(sizeof(*data__files) * data__files_cnt);
1911         if (!data__files)
1912                 return -ENOMEM;
1913
1914         data__for_each_file(i, d) {
1915                 struct perf_data *data = &d->data;
1916
1917                 data->path  = use_default ? defaults[i] : argv[i];
1918                 data->mode  = PERF_DATA_MODE_READ;
1919                 data->force = force;
1920
1921                 d->idx  = i;
1922         }
1923
1924         return 0;
1925 }
1926
1927 static int diff__config(const char *var, const char *value,
1928                         void *cb __maybe_unused)
1929 {
1930         if (!strcmp(var, "diff.order")) {
1931                 int ret;
1932                 if (perf_config_int(&ret, var, value) < 0)
1933                         return -1;
1934                 sort_compute = ret;
1935                 return 0;
1936         }
1937         if (!strcmp(var, "diff.compute")) {
1938                 if (!strcmp(value, "delta")) {
1939                         compute = COMPUTE_DELTA;
1940                 } else if (!strcmp(value, "delta-abs")) {
1941                         compute = COMPUTE_DELTA_ABS;
1942                 } else if (!strcmp(value, "ratio")) {
1943                         compute = COMPUTE_RATIO;
1944                 } else if (!strcmp(value, "wdiff")) {
1945                         compute = COMPUTE_WEIGHTED_DIFF;
1946                 } else {
1947                         pr_err("Invalid compute method: %s\n", value);
1948                         return -1;
1949                 }
1950         }
1951
1952         return 0;
1953 }
1954
1955 int cmd_diff(int argc, const char **argv)
1956 {
1957         int ret = hists__init();
1958
1959         if (ret < 0)
1960                 return ret;
1961
1962         perf_config(diff__config, NULL);
1963
1964         argc = parse_options(argc, argv, options, diff_usage, 0);
1965
1966         if (quiet)
1967                 perf_quiet_option();
1968
1969         if (cycles_hist && (compute != COMPUTE_CYCLES))
1970                 usage_with_options(diff_usage, options);
1971
1972         if (pdiff.stream)
1973                 compute = COMPUTE_STREAM;
1974
1975         symbol__annotation_init();
1976
1977         if (symbol__init(NULL) < 0)
1978                 return -1;
1979
1980         if (data_init(argc, argv) < 0)
1981                 return -1;
1982
1983         if (check_file_brstack() < 0)
1984                 return -1;
1985
1986         if ((compute == COMPUTE_CYCLES || compute == COMPUTE_STREAM)
1987             && !pdiff.has_br_stack) {
1988                 return -1;
1989         }
1990
1991         if (compute == COMPUTE_STREAM) {
1992                 symbol_conf.show_branchflag_count = true;
1993                 symbol_conf.disable_add2line_warn = true;
1994                 callchain_param.mode = CHAIN_FLAT;
1995                 callchain_param.key = CCKEY_SRCLINE;
1996                 callchain_param.branch_callstack = 1;
1997                 symbol_conf.use_callchain = true;
1998                 callchain_register_param(&callchain_param);
1999                 sort_order = "srcline,symbol,dso";
2000         } else {
2001                 if (ui_init() < 0)
2002                         return -1;
2003
2004                 sort__mode = SORT_MODE__DIFF;
2005         }
2006
2007         if (setup_sorting(NULL) < 0)
2008                 usage_with_options(diff_usage, options);
2009
2010         setup_pager();
2011
2012         sort__setup_elide(NULL);
2013
2014         return __cmd_diff();
2015 }